openmp tips, tricks and gotchas...parallel regions • the overhead of executing a parallel region...
TRANSCRIPT
OPENMP TIPS, TRICKS AND GOTCHAS
Mark Bull EPCC, University of Edinburgh (and OpenMP ARB) [email protected]
Directives
• Mistyping the sentinel (e.g. !OMP or #pragma opm ) typically raises no error message.
• Be careful!
• Extra nasty if it is e.g. #pragma opm atomic – race condition!
• Write a script to search your code for your common typos
OpenMPCon 2015 2
Writing code that works without OpenMP too
• The macro _OPENMP is defined if code is compiled with the OpenMP switch. • You can use this to conditionally compile code so that it works with
and without OpenMP enabled.
• If you want to link dummy OpenMP library routines into sequential code, there is code in the standard you can copy (Appendix A in 4.0)
OpenMPCon 2015 3
Parallel regions • The overhead of executing a parallel region is typically in the
tens of microseconds range • depends on compiler, hardware, no. of threads
• The sequential execution time of a section of code has to be several times this to make it worthwhile parallelising.
• If a code section is only sometimes long enough, use the if clause to decide at runtime whether to go parallel or not. • Overhead on one thread is typically much smaller (<1µs).
• You can use the EPCC OpenMP microbenchmarks to do detailed measurements of overheads on your system.
• Download from www.epcc.ed.ac.uk/research/computing/performance-characterisation-and-benchmarking
OpenMPCon 2015 4
Is my loop parallelisable? • Quick and dirty test for whether the iterations of a loop are
independent. • Run the loop in reverse order!! • Not infallible, but counterexamples are quite hard to construct.
OpenMPCon 2015 5
Loops and nowait #pragma omp parallel { #pragma omp for schedule(static) nowait for(i=0;i<N;i++){ a[i] = .... } #pragma omp for schedule(static) for(i=0;i<N;i++){ ... = a[i] } }
• This is safe so long as the number of iterations in the two loops and the schedules are the same (must be static, but you can specify a chunksize)
• Guaranteed to get same mapping of iterations to
threads.
OpenMPCon 2015 6
Default schedule • Note that the default schedule for loops with no schedule
clause is implementation defined. • Doesn’t have to be STATIC. • In practice, in all implementations I know of, it is. • Nevertheless you should not rely on this! • Also note that SCHEDULE(STATIC) does not completely
specify the distribution of loop iterations. • don’t write code that relies on a particular mapping of iterations to
threads
OpenMPCon 2015 7
Tuning the chunksize • Tuning the chunksize for static or dynamic schedules can be
tricky because the optimal chunksize can depend quite strongly on the number of threads.
• It’s often more robust to tune the number of chunks per thread and derive the chunksize from that. • chunksize expression does not have to be a compile-time constant
OpenMPCon 2015 8
SINGLE or MASTER? • Both constructs cause a code block to be executed by one
thread only, while the others skip it: which should you use?
• MASTER has lower overhead (it’s just a test, whereas SINGLE requires some synchronisation).
• But beware that MASTER has no implied barrier!
• If you expect some threads to arrive before others, use SINGLE, otherwise use MASTER
OpenMPCon 2015 9
Data sharing attributes • Don’t forget that private variables are uninitialised on entry to
parallel regions!
• Can use firstprivate, but it’s more likely to be an error. • use cases for firstprivate are surprisingly rare.
OpenMPCon 2015 10
Default(none) • The default behaviour for parallel regions and worksharing
construct is default(shared)
• This is extremely dangerous - makes it far too easily to accidentally share variables.
• Possibly the worst design decision in the history of OpenMP!
• Always, always use default(none) • I mean always. No exceptions! • Everybody suffers from “variable blindness”.
OpenMPCon 2015 11
Spot the bug! #pragma omp parallel for private(temp) for(i=0;i<N;i++){
for (j=0;j<M;j++){ temp = b[i]*c[j]; a[i][j] = temp * temp + d[i]; } }
• May always get the right result with sufficient compiler optimisation!
OpenMPCon 2015 12
Private global variables double foo; #pragma omp parallel \ private(foo) { foo = .... a = somefunc(); }
extern double foo;
double sumfunc(void){
... = foo;
}
• Unspecified whether the reference to foo in somefunc is to the original storage or the private copy.
• Unportable and therefore unusable! • If you want access to the private copy, pass it through the
argument list (or use threadprivate).
OpenMPCon 2015 13
Huge long loops • What should I do in this situation? (typical old-fashioned
Fortran style)
do i=1,n ..... several pages of code referencing 100+ variables end do
• Determining the correct scope (private/shared/reduction) for all those variables is tedious, error prone and difficult to test adequately.
OpenMPCon 2015 14
• Refactor sequential code to do i=1,n call loopbody(......) end do
• Make all loop temporary variables local to loopbody • Pass the rest through argument list • Much easier to test for correctness! • Then parallelise...... • C/C++ programmers can declare temporaries in the scope of
the loop body.
OpenMPCon 2015 15
Reduction race trap #pragma omp parallel shared(sum, b) { sum = 0.0; #pragma omp for reduction(+:sum) for(i=0;i<n:i++) { sum += b[i]; } .... = sum; }
• There is a race between the initialisation of sum and the updates to it at the end of the loop.
OpenMPCon 2015 16
Missing SAVE or static • Compiling my sequential code with the OpenMP flag caused it
to break: what happened? • You may have a bug in your code which is assuming that the
contents of a local variable are preserved between function calls. • compiling with OpenMP flag forces all local variables to be stack
allocated and not heap allocated • might also cause stack overflow
• Need to use SAVE or static correctly • but these variables are then shared by default • may need to make them threadprivate • “first time through” code may need refactoring (e.g. execute it before the
parallel region)
OpenMPCon 2015 17
Stack size • If you have large private data structures, it is possible to run
out of stack space. • The size of thread stack apart from the master thread can be
controlled by the OMP_STACKSIZE environment variable. • The size of the master thread’s stack is controlled in the same
way as for sequential program (e.g. compiler switch or using ulimit ). • OpenMP can’t control this as by the time the runtime is called it’s too
late!
OpenMPCon 2015 18
Critical and atomic • You can’t protect updates to shared variables in one place
with atomic and another with critical, if they might contend. • No mutual exclusion between these
• critical protects code, atomic protects memory locations. #pragma omp parallel { #pragma omp critical a+=2; #pragma omp atomic a+=3; }
OpenMPCon 2015 19
Allocating storage based on number of threads • Sometimes you want to allocate some storage whose size is
determined by the number of threads. • but how do you know how many threads the next parallel region will
use?
• Can call omp_get_max_threads() which returns the value of the nthreads-var ICV. The number of threads used for the next parallel region will not exceed this • except if a num_threads clause is used.
• Note that the implementation can always deliver fewer threads than this value • if your code depends on there actually being a certain number of
threads, you should always call omp_get_num_threads() to check
OpenMPCon 2015 20
Environment for performance • There are some environment variables you should set to
maximise performance. • don’t rely on the defaults for these!
OMP_WAIT_POLICY=active • Encourages idle threads to spin rather than sleep OMP_DYNAMIC=false • Don’t let the runtime deliver fewer threads than you asked for OMP_PROC_BIND=true • Prevents threads migrating between cores
OpenMPCon 2015 21
Debugging tools • Traditional debuggers such as DDT or Totalview have support
for OpenMP
• This is good, but they are not much help for tracking down race conditions • debugger changes the timing of event on different threads
• Race detection tools work in a different way • capture all the memory accesses during a run, then analyse this data for
races which might have occured. • Intel Inspector XE • Oracle Solaris Studio Thread Analyzer
OpenMPCon 2015 22
Timers • Make sure your timer actually does measure wall clock time!
• Do use omp_get_wtime() !
• Don’t use clock() for example • measures CPU time accumulated across all threads • no wonder you don’t see any speedup......
OpenMPCon 2015 23